Groundbreaking neuroscience research reveals that the human brain processes spoken language using a stepwise, hierarchical procedure remarkably similar to advanced AI models like GPT. A study led by scientists from The Hebrew University of Jerusalem and Princeton University, published in Nature Communications, found that deeper layers of a large language model correspond to later stages of brain activity during comprehension.
Researchers used electrocorticography (ECoG) to record neural activity from participants listening to a 30-minute narrative podcast. They discovered that early brain responses align with the shallow processing layers of an AI model, which handle basic word features. In contrast, later neural activity in high-level language regions, particularly Broca's area, aligns with the AI's deeper, contextual layers that build meaning. This suggests both systems converge on a similar step-by-step path to understanding.
The findings challenge long-standing, rule-based theories of language comprehension, supporting a more dynamic, statistical model where meaning emerges from context. The research team has publicly released the complete neural dataset, providing a new resource for scientists worldwide to test theories of language processing.